昨天为公司多个部门的小伙伴们做了关于TiDB的科普，把写的讲义直接贴在下面吧。

（Markdown直接渲染成网页真的比做PPT简便太多了

---

TiDB 101

---

Part I - Introduction

What is TiDB?

• TiDB is an open-source, distributed, relational database
• It is NewSQL -- taking advantage of both traditional RDBMS and NoSQL
• With native OLTP and optional OLAP workload support
• With high compatibility against MySQL ecosystem

Key Features

• Horizontally scaling out/scaling in flexibly [Computing & storage separation]
• High availability [Multi-Raft consensus]
• ACID guarantee [Percolator distributed transaction]
• HTAP ready [Row-based engine for OLTP, column-based engine for OLAP]
• Rich series of data migration tools for transferring data
• 90% compatible with MySQL 5.7 environment
  • See https://docs.pingcap.com/tidb/stable/mysql-compatibility for compatibility issues
  • See https://docs.pingcap.com/tidb/stable/tidb-limitations for limits

Typical Use Cases

• Financial industry scenarios with high requirements for consistency, reliability, and fault tolerance
• Massive data and high concurrency scenarios with high requirements for capacity, scalability, and concurrency
• Scattered data collection and secondary processing scenarios

Disadvantages?

• ...

---

Part II - Architecture Overview

TiKV - Storage

• Storage layer for OLTP
• A distributed transactional key-value engine
• TiKV stores are built on top of RocksDB (LSMT-based local-embedded KV system)
• Each store is divided into regions -- the basic unit of organizing data
  • A region contains a continuous, sorted key range of [start_key, end_key), like an SST
  • Regions are maintained in multiple replicas (default 3) as Raft groups, having a cap size, and can be merged & split
  • Raft provides safe leader election / log replication / membership changing

• Table data are mapped to TiKV with rules
  • Row: t{tableID}_r{rowID} -> [colValue1, colValue2, colValue3, ...]
  • Unique index: t{tableID}_i{indexID}_{colValue} -> [rowID]
  • Ordinary index: t{tableID}_i{indexID}_{colValue}_{rowID} -> [null]
• Single-row (Per-KV) transaction is natively supported with snapshot isolation (SI) level
  • Distributed transaction enabled by implementing Percolator protocol
  • Almost decentralized 2PC
  • MVCC timestamps stored in different RocksDB column families

TiDB - Computation

• Stateless SQL layer
• Expose MySQL-flavored endpoint to client connections
  • Can provide unified interface through load balancing components (HAProxy, LVS, Aliyun SLB, etc.)
• Perform SQL parsing & optimization
• Generate distributed execution plan
• Execute through DistSQL API (with TiKV coprocessors computing at lower level) or raw KV API
• Transmit data request/response to/from storage nodes
• e.g. A simple query showing TiDB & TiKV working together

PD ("Placement Driver") - Scheduling

• Hold metadata (data location) and schedule regions in the cluster
• Receive heartbeats (state information) from TiKV peers and region leaders
• Act as timestamp oracle (TSO) for transactions
• Provide guarantee for following strategies
  • The number of replicas of a Region needs to be correct
  • Replicas of a Region need to be at different positions
  • Replicas and their leaders need to be balanced between stores
  • Hot-spots and storage size need to be balanced between stores
• e.g. Region split & merge

• e.g. Hot-spot removal

TiSpark - OLAP Extension [Optional]

• Traditional OLAP solution
• Shim layer built for runnning Spark on TiKV cluster
• Taking advantage of Spark SQL Catalyst

TiFlash - OLAP Extension [Optional]

• Brand-new OLAP solution
• Columnar storage with a layer of coprocessors efficiently implemented by ClickHouse
• Data are replicated from TiKV to TiFlash asynchronously as Raft learner
• Somewhat expensive, and we've already got a real ClickHouse cluster running for 8 months =.=

---

Part III - Peripheral Tools

TiUP

• A command line tool that manages components in the TiDB cluster
• Deploy / config / start / stop / scale / rolling upgrade with simple commands and YAML files

TiCDC

• A tool for replicating incremental data (change data capture) of TiDB
• Support other downstream systems to subscribe to TiDB binlog

DM (Data Migration)

• An integrated data migration task platform that supports full/incremental data syncing from MySQL/MariaDB into TiDB
• Master-worker style
• Workers act as MySQL slaves to keep track of binlog from sources
• Supports table routing, black/white list, binlog filter & online DDLs
• Supports merging from sharded databases/tables

• Our proposed shard merging & CDC tracking route by now

---

Part IV - Cluster Configuration

• All servers are Aliyun ECSs (type g6e), CentOS 7.3+, 10Gbps network interfaces
• Regular parameter tuning for big data servers (THP, swap, CPU policy, NTP, etc...)
• Use TiUP for deployment
• See https://docs.pingcap.com/tidb/stable/hardware-and-software-requirements for detailed requirements

TiDB & PD

• TiDB & PD can be deployed on the same server
• TiDB requires CPU & memory while PD requires fast I/O
• Minimum 3 instances
• 3 * ecs.g6e.8xlarge [32vCPU / 128GB] & 2 * 1TB SSD

TiKV

• Multiple TiKVs can be deployed on the same server server
• Requires CPU & memory with very fast I/O
• Minimum 3 instances (Raft requires at least n / 2 + 1 nodes functioning)
• 6 * ecs.g6e.13xlarge [52vCPU / 192GB] & 3 * 2TB ESSD PL2
• 3TiKVs per server = 18 instances total
  • See https://docs.pingcap.com/tidb/stable/hybrid-deployment-topology#key-parameters for crucial parameters
  • Assign a different disk to each TiKV

TiCDC

• Minimum 2 instances
• 3 * ecs.g6e.4xlarge [16vCPU / 64GB] & 1TB SSD

DM

• DM master & worker can be deployed on the same server
• 3 * ecs.g6e.4xlarge [16vCPU / 64GB] & 1TB HDD
• When dumping data, make sure that the disk has enough free space to hold them

Monitoring

• Including Prometheus, Grafana & AlertManager
• 1 * ecs.g6e.4xlarge [16vCPU / 64GB] & 500GB SSD

---

Part V - Operation & Monitoring

TiDB Dashboard

• Overview

• Query summary

【此处涉及机密数据，故略去】

• Query details

【此处涉及机密数据，故略去】

• Key visualization

【此处涉及机密数据，故略去】

Grafana Dashboards (w/ Prometheus)

Alerting (by Mail)

---

Part VI - Examples, Q&A

• Client connection
• Dashboards
• Monitoring
• DM tasks
• ...

---

The End